performance tuning

[David Xu]

Julian Elischer wrote:

>
>
> Julian Elischer wrote:
>
>>
>>
>> Petri Helenius wrote:
>>
>>>
>>>  Now that 5.3 is the first STABLE release, is the future focus on
>>>  -threads towards performance, so that benchmarks and profiling, etc.
>>>  would be appreciated somewhat more than they might have been
>>>  previously when the development was more in the "just get it working
>>>  right" mode?
>>>
>> Basically, yes.
>> there is still work to do in the "get it right" field but
>> it's working "right enough" now that performance is
>> starting to be an issue.
>>
>> There is work proceeding on 3 fronts.
>>
>> * libthr continues to be improved by Mike.
>>
>> * libpthread is being scrutinised for ineffiencies and problems.
>>   In particular M:N performance has a lot to be desired.
>>
>> * (from left field) David has been working on a hybrid of libpthread and
>>  libthr.  As a private project, David is playing w 
>
>
> oops accidentally deleted some of my own mail..
>
>
>
> -- with a version of libpthread optimised for 1:1 using libthr 
> entrypoints.
>
You can try:
http://people.freebsd.org/~davidxu/kern_thr.c
http://people.freebsd.org/~davidxu/kern_umtx.c
http://people.freebsd.org/~davidxu/libfptl.tgz
I can run all applications current libpthread can run.
Performance still needs to be tuned in kern_umtx.c, but it
already beats current libpthread performance using supersmacks
with 100 clients.   :-)

David Xu